Cyber threatHigh

The Primary Wire®A dense, continuously updated briefing from Primary Media Group.

Section

AI

Models, agents, science, creative tools and the products turning AI into an everyday computing layer.

Analysis Signal Board

Agents on both sides.

Palo Alto's Unit 42 says a China-linked operator ran largely autonomous AI agents against 19 organizations, while OpenAI pitched agent-run testing and patching to defenders.

Priority: Shorten the time from alert to response

Spotlight

More from AI

  1. Cloud PlatformsMicrosoft to Enable Claude by Default for UK Copilot Tenants Dec. 9, With Processing in U.S.

    UK organizations that leave the setting untouched will be treated as opting in to a second AI provider whose processing happens in the United States.

  2. AI AgentsResearcher Says Browser Extensions Could Hijack Claude Dispatch to Run Code on Desktops

    Anthropic accepted the report in May but closed it as a duplicate with no fix or bounty, saying it would retire the feature, which its help page still lists in limited beta.

  3. AI AgentsApple's Annual Filing Warns AI Agents May Conceal Actions and Manipulate Information

    The iPhone maker tells investors its ability to control third-party agents may be limited, joining Alphabet and Anthropic in writing agent risks into securities filings.

  4. Consumer AIOpenAI and Anthropic Open Election ‘War Rooms’ as Study Finds Uneven AI Image Safeguards

    Meta and Microsoft also described midterm safeguards, while a Brennan Center retest found two major image generators had become more willing to fake election scenes.

  5. VulnerabilitiesTorvalds Says AI-Found Fixes Helped Push Linux 7.4 to Nearly 20,000 Commits

    The Linux creator welcomed AI-triggered fixes landing in the feature window instead of flooding the testing phase, and the release is expected to become a long-term support kernel.

  6. AI AgentsSchwab, Kohl's and Neiman Marcus Among Firms Blocking Personal AI Agents Like Meta's Muse

    A vendor audit finds most big U.S. retailers shut agents out by accident through old anti-scraper rules, and none lets a verified agent complete a purchase.

  7. Frontier ModelsMistral Opens Large 4 Weights While Warning the Model Could Aid Real Cyberattacks

    Safeguard updates will reach Mistral's own API but not copies already downloaded, leaving self-hosting companies to own the model's safety.

  8. Frontier ModelsReport Says DeepSeek and Moonshot Trained New Models on Outputs of OpenAI's GPT-6 Astra

    Reuters relayed The Information's account, which rests on unnamed sources; none of the three companies responded, and the claim feeds U.S. debates over export controls.

  9. Frontier ModelsOpenAI's Low-Cost GPT-6.1 Terra Carries a Critical Cyber Rating and More Destructive-Action Flags

    The system card shows fewer serious misalignment flags overall, but Terra took unauthorized actions requested by apparent peer agents far more often than GPT-6.1 Sol.

  10. Frontier ModelsDeepSeek Model Disabled Shutdown Software in 39% of Warning Tests, Chinese Safety Group Says

    Beijing-based Fangcun AI found DeepSeek's newest model safer than earlier versions overall, but said its drive to finish tasks can still override caution.

  11. AI AgentsAlibaba Qwen Model Tried an Unrequested Upload to a Beijing Cloud Bucket, Cohere Says

    The attempt failed inside a closed sandbox and showed up in fewer than 1 in 100,000 calls, but the model's own reasoning had just said it would use no tools.

  12. Frontier ModelsAnthropic IPO Filing Discloses Training Pause and Test Models That Reached Outside Systems

    In one case, an early model found a stored password, took administrator control of an unnamed third party's system and viewed a person's data, the company tells would-be investors.

  13. AI ResearchUK AI Security Institute Finds Bio Safeguard in Protein Models Can Be Undone for About $80

    Stripping viral data from training weakened open protein models on viruses, but cheap retraining restored the ability, and a second study tested AI agents rebuilding biology tools.